{"id":"97ddb722-91ca-4b55-9720-29f7f1147bad","arxiv_id":"2412.07616","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PVP pushes 3D occupancy prediction accuracy on the OpenOccupancy benchmark by combining polar voxel grids with a global attention module and decomposed convolutions.","lead":"The paper introduces PVP, a neural network that predicts 3D semantic occupancy around a self-driving car using a polar coordinate grid instead of a Cartesian one. It reports large accuracy gains on the OpenOccupancy benchmark from two new modules that correct distortions and spread features correctly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed SOTA margins in Table 1 rest entirely on baseline numbers taken from prior work; without evidence that PointOcc and M-CONet were run under PVP's exact protocol, the reported 1.9/2.9 and 7.9/6.8 deltas are unsupported.","rationale":"The strongest claim is the state-of-the-art result. The necessary condition for that claim is that the comparison against prior methods is fair. The paper gives only a one-sentence assertion and takes all baseline numbers from the literature/benchmark. Since the method uses a non-Cartesian grid and a custom sampling head, even the evaluation pipeline could differ. The paper's own ablation shows that polar representation alone is not beneficial; therefore the reported SOTA gains are produced by the novel modules on top of polar representation. Without rerunning baselines, there is no way to know if the baselines were trained or evaluated in the same regime. This is not a dispute about the internal logic; the ablations are consistent. It is an external-validity gap. The concrete test is straightforward and would settle it. The reader's verdict of CONDITIONAL is appropriate: the paper should be accepted only if the authors release code and reproduce the comparison, or the claims should be downgraded. I see no reason to change the verdict to REJECT because the method could be correct; the evidence is just incomplete.","tokens_in":13787,"tokens_out":7460,"duration_ms":60856,"concrete_test":"Obtain the official PointOcc (LiDAR) and M-CONet (C&L) checkpoints and evaluate them with the exact OpenOccupancy evaluation script used for PVP, including the same Cartesian 512x512x40 grid and the same polar-to-Cartesian output conversion applied to PVP's predictions. Verify that the reproduced IoU/mIoU match Table 1 (PointOcc: 34.1/23.9; M-CONet: 29.5/20.1). Then retrain both baselines under PVP's 25-epoch, batch-size-8, Adam/cosine schedule with the same loss weights and data split, and recompute the deltas. If the baselines reproduce exactly and the margins persist, the central claim stands; if the reproduced numbers differ, the reported gains must be revised to reflect the matched-protocol comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: PVP surpasses PointOcc by +1.9 mIoU/+2.9 IoU and M-CONet by +7.9 mIoU/+6.8 IoU on OpenOccupancy (Table 1). Section 4.1.2 asserts that 'All baseline implementations were aligned with the benchmark [16] to ensure fairness in comparison,' but the paper reruns no baseline, provides no training hyper-parameters for the baselines, and offers no logs or evaluation scripts. Small discrepancies in training epochs, learning-rate decay, voxelization range, loss weights, or in how the polar-grid prediction is sampled to the 512x512x40 Cartesian evaluation volume can shift these numbers by more than a point. Table 2 makes this concern acute: switching from Cartesian to polar alone reduces IoU by 0.4% (26.4 to 26.0) at 5 epochs, so all of the claimed superiority comes from the GRP+PD-Conv components under a specific training schedule. Unless the baselines were reproduced under that same schedule, the SOTA claim cannot be separated from protocol mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PVP, a 3D semantic occupancy predictor that operates in polar coordinates rather than the usual Cartesian volume. It introduces two components: a Global Represent Propagation (GRP) module that condenses local windows and then performs axis-wise global attention to propagate features, and a Plane Decomposed Convolution (PD-Conv) that replaces 3D convolutions with three 2D convolutions on different planes. The model is built on the CONet/OpenOccupancy pipeline, loss, and prediction head. On the OpenOccupancy (nuScenes) validation set, the authors report that LiDAR-only PVP exceeds PointOcc by 1.9 mIoU and 2.9 IoU, and that multimodal PVP exceeds M-CONet by 7.9 mIoU and 6.8 IoU. Component ablations at 5 epochs show consistent gains for GRP and PD-Conv over a polar-only baseline. The paper does not release code, logs, or rerun baseline methods; all comparisons rely on published numbers.","tokens_in":14037,"tokens_out":5834,"duration_ms":53846,"significance":"If the reported empirical gains are reproducible under a fair protocol, the polar-coordinate design is a meaningful step for 3D semantic occupancy prediction, and the two proposed modules address a real distortion problem. The paper's internal ablations are systematic and show that the full model is much stronger than a direct polar conversion, which is a useful finding. However, the central SOTA claim rests entirely on baseline numbers taken from prior papers without any evidence that training schedules, voxelization ranges, loss weights, or evaluation code match. The absence of error bars, multi-seed runs, and code further weakens the quantitative claims. The contribution is therefore defensible in design but not yet supported in its headline comparison.","major_comments":[{"comment":"The paper's central claim—surpassing PointOcc by +1.9 mIoU/+2.9 IoU and M-CONet by +7.9 mIoU/+6.8 IoU—is supported only by published baseline numbers. The sentence in Section 4.1.2, 'All baseline implementations were aligned with the benchmark [16] to ensure fairness in comparison,' is not backed by any rerun, training log, configuration file, or evaluation script. Small protocol differences (number of epochs, learning-rate schedule, voxelization range, polar-to-Cartesian sampling, loss weighting) can plausibly shift mIoU by more than a point. Since Table 1 is the only evidence for the SOTA claim, the authors need to provide either rerun baselines under the exact PVP protocol or explicitly state that the compared numbers are quoted from the original papers and soften the 'surpasses' wording accordingly.","section":"4.1.2, Table 1"},{"comment":"All ablations are single-run results at 5 epochs, while the headline results in Table 1 are at 25 epochs. There is no evidence that the 5-epoch ranking is representative of the 25-epoch setting. In Table 3, four variants (Channel, Swin, GRR, and Polar-no-attention) are separated by only 0.1 mIoU, which is almost certainly within run-to-run noise. The claims that GRP is better than these alternatives, and that the PD-Conv structure variations are negligible, are therefore underdetermined. The authors should provide error bars or multiple seeds for the key rows, or run the full 25-epoch protocol at least for the main ablation path.","section":"4.4, Tables 2, 3, 5"},{"comment":"The GRP module is a core contribution, but Eq. (2) is not a well-formed attention operation. It is written as `softmax( Q_i K_i / sqrt(d) · V_i + E(p) )`, which appears to multiply the attention weights by V without summing over key positions, and it adds positional encoding after the dot product rather than inside the softmax. The `maxsel` operation is not formally defined, the relation between the window size `Wg` and the downsampling factor `S` in the condensed feature map is never stated, and the reverse cross attention in Section 3.2.2 uses an undefined dimension `N`. As written, this part of the method cannot be reproduced or checked.","section":"3.2.1, Eq. (2); 3.2.2"},{"comment":"PD-Conv, the second core contribution, is under-specified. The text says the three decomposed blocks are stacked in series and then combined in parallel, but it does not give the exact block order, the number of blocks, kernel sizes, strides, padding, how the parallel branches are aggregated, or which stages of the 3D backbone are replaced. The statement that 'the stacking order does not significantly affect performance' is qualitative and is only validated at 5 epochs. Without a precise specification, the method is not reproducible from the paper, and the claimed source of the gain (plane decomposition rather than asymmetric kernel layout) cannot be independently verified.","section":"3.3, 4.2"}],"minor_comments":[{"comment":"The rows labeled 'OpenOccupancy' report IoU/mIoU of 30.7/15.0 (L) and 29.3/19.8 (C&L), while Table 1 lists L-CONet as 30.9/15.8 and M-CONet as 29.5/20.1. Please clarify whether 'OpenOccupancy' and 'CONet' refer to the same implementation and why the numbers differ; otherwise the efficiency comparison is ambiguous.","section":"Table 4"},{"comment":"The abstract says 'LiDAR-only methods' and 'multimodal methods' without naming the specific baselines; the actual comparison is against PointOcc and M-CONet. Please name the strongest prior method in each category to make the claim precise.","section":"Abstract, Section 4.3"},{"comment":"The dataset description says '32-lane LiDAR'; this should be '32-beam LiDAR'.","section":"4.1.1"},{"comment":"There is a likely typo in Eq. (2): 'sof tmax' should be 'softmax', and the inner expression lacks the conventional scaling/summation notation. Please rewrite the equation in standard form.","section":"3.2.1"},{"comment":"The Figure 3 caption is incomplete: it ends without a period and contains 'The pipeline ofPVP'. Please proofread the captions and the main text for similar typographical issues (e.g., 'tenser', '3D V oxel-based').","section":"Figure 3"},{"comment":"The list of compared baselines does not mention the training schedule or evaluation protocol used for each published baseline. Even if the numbers are quoted, a sentence stating that no baseline was rerun would help readers assess the fairness claim.","section":"Section 4.1.2"}],"recommendation":"major_revision","confidential_remarks":"The main issue is verifiability of the SOTA claim. The authors should be asked to provide code, logs, or a documented protocol for comparing with PointOcc and M-CONet. If that is not feasible, the paper should be revised to present the results as a comparison with published numbers under a shared pipeline, not as definitive 'surpasses' evidence. The absence of multi-seed results is particularly problematic for the 0.1 mIoU differences used to dismiss alternative attention modules."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper that claims a new SOTA on OpenOccupancy with a polar voxel predictor. The design is a sensible combination of existing ideas — polar/cylindrical grids, asymmetric convolutions from Cylinder3D, axial attention — plus a Global Representation Propagation module and a Plane-Decomposed Convolution. The ablations are internally consistent and show each component helps, which is the paper's main strength.\n\nWhat's genuinely new: the application of a dense polar volume to occupancy prediction with an explicit attention module for long-range feature propagation, and a decomposed convolution that treats range/azimuth/height distortions separately. The efficiency analysis is also useful: PVP gets its gains at roughly the same cost as CONet.\n\nSoft spots: the central SOTA claims (1.9/2.9 over PointOcc, 7.9/6.8 over M-CONet) rely entirely on published baseline numbers. The paper does not rerun any baseline, provides no code, no training logs, and no error bars. Section 4.1.2's assertion that 'all baseline implementations were aligned with the benchmark' does not substitute for reproducibility. The stress-test note is right that Table 2 makes this acute: going Cartesian-to-polar by itself loses 0.4 IoU, so the reported advantages come entirely from the new modules under a specific training schedule. A small protocol mismatch in voxelization range, loss weights, or sampling to the Cartesian evaluation volume could shift these numbers by more than a point. I wouldn't call this fraud, but the magnitude of the claim currently has an unsupported precision. The write-up also leaves a few implementation details underspecified (window size for local condense attention, the downsampling factor S, and the exact reverse cross-attention), and there are minor typos.\n\nOverall: the paper deserves a serious referee. It's a legitimate architectural contribution to a benchmark that many labs use, and the ablations suggest the gains are real. But the published version needs code and ideally baseline reruns, or at minimum a clear statement of which hyperparameters were matched. I'd send it to peer review with a request for reproducibility materials.","headline":"Solid incremental architecture with strong claimed gains on OpenOccupancy, but the SOTA margins rest on unverified baseline numbers and no code; worth refereeing but require reproducibility materials.","tokens_in":14581,"tokens_out":2062,"would_cite":true,"duration_ms":17835,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Polar voxel grids, with distortion-correcting modules, outdo Cartesian methods for 3D occupancy prediction.","keywords":["3D semantic occupancy prediction","polar representation","feature distortion","attention mechanism","plane decomposed convolution","LiDAR-camera fusion","autonomous driving"],"falsifier":"Rerun PointOcc and M-CONet with the exact training schedule (25 epochs, batch size 8, V100 GPUs), voxel range [0.3, 73]m on the rho axis, and the same evaluation code; if the mIoU gaps over these baselines shrink to well below the reported 1.9% and 7.9%, the central claim of state-of-the-art performance is refuted.","tokens_in":13570,"feed_emoji":"🧭","tokens_out":6872,"duration_ms":55336,"temperature":0.7,"pith_summary":"This paper argues that representing the surrounding 3D space in polar coordinates, with finer voxels near the sensor and coarser voxels far away, can beat the standard Cartesian voxel grid for semantic occupancy prediction provided the distortions polar grids introduce are explicitly handled. To that end it proposes PVP, built from two modules: GRP, an attention mechanism that propagates features along the radial, azimuth, and height axes so large structures like roads stay coherent, and PD-Conv, which replaces full 3D convolutions with three plane-wise 2D convolutions that match how distortion actually varies across views. On the OpenOccupancy benchmark, PVP reports a 1.9% mIoU and 2.9% IoU gain over the best LiDAR-only method and 7.9% mIoU and 6.8% IoU over the best multimodal baseline. A sympathetic reader would take the paper's central claim to be that polar representation, once its distortions are corrected, is not merely viable but superior for dense 3D scene understanding.","feed_headline":"Polar voxel model tops 3D occupancy benchmarks","feed_subtitle":"PVP beats LiDAR-only and multimodal baselines on OpenOccupancy, lifting mIoU by up to 7.9 points.","key_machinery":"The load-bearing machinery is the pair of distortion-countering modules. GRP first applies local condense attention, which selects a representative feature per non-overlapping window and refines it with cross-attention against the window's own features, producing a condensed volume; then global decomposed attention runs self-attention in strips along the radial, azimuth, and height axes in turn, so distant parts of a road or building can interact at low cost, and finally a reverse cross-attention propagates the calibrated features back to the full resolution. PD-Conv replaces each 3x3x3 convolution with three stacked decomposed convolutions (1x3x3, 3x1x3, 3x3x1), each acting on one plane of the polar volume, which separates the scale, projection, and identity distortions onto different planes. The final output is resampled to Cartesian voxels via trilinear sampling and passed through the same occupancy head, losses, and coarse-to-fine refinement as the OpenOccupancy baseline.","core_discovery":"PVP is a 3D semantic occupancy predictor that operates on a polar voxel grid instead of a Cartesian one. The paper's central claim is that the two design elements it introduces resolve the two failure modes that have held polar representations back in dense prediction: GRP (Global Representation Propagation) fixes the global misalignment of large 'stuff' classes by condensing local windows into representative features, running decomposed attention along each axis, and propagating the corrected features back to every voxel; PD-Conv (Plane Decomposed Convolution) fixes local distortion by replacing each 3D convolution with three 2D convolutions, a scale transformation on the range plane, a projection transformation on the BEV plane, and an identity transformation on the slicing plane. With these changes, PVP reports the strongest results on the OpenOccupancy validation set, improving IoU to 37.0% (LiDAR-only) and 36.3% (camera+LiDAR) and mIoU to 25.8% and 28.0%, respectively, surpassing the previous state of the art in each input modality.","pith_inferences":["Because the baselines are quoted from published results rather than re-run, the exact margins of 1.9/2.9 and 7.9/6.8 percentage points should be treated as conditional on protocol alignment until an independent re-implementation confirms them.","The observation that reordering the three decomposed convolutions barely changes performance hints that the benefit comes from separating the axes, not from the specific receptive field; a testable extension is to use randomly permuted kernel orders and measure the variance.","GRP's design could transfer to other dense prediction tasks dominated by large background classes (e.g., BEV map segmentation), where global feature propagation along structured axes may repair similar distortion failures."],"forward_implications":["If the reported numbers hold, polar-coordinate voxelization becomes the recommended representation for 3D semantic occupancy prediction in autonomous driving, because it spends resolution where sensors actually provide information.","PVP's gains are largest on stuff classes such as roads, sidewalks, and terrain, suggesting the approach specifically repairs the failure mode that Cartesian methods share for large, elongated structures.","The method keeps memory use near or below the Cartesian baseline while improving accuracy, per the paper's efficiency analysis, making it viable for onboard deployment.","The polar-to-Cartesian trilinear sampling at the output head means downstream consumers of the predictions still receive a standard Cartesian voxel grid."],"supporting_citations":[{"why":"Supplies the OpenOccupancy dataset, evaluation protocol, and the CONet baselines (L-CONet, M-CONet, C-CONet) that PVP inherits its pipeline from and compares against.","marker":"[16]"},{"why":"The prior polar-based LiDAR-only method that PVP claims to surpass by 1.9% mIoU and 2.9% IoU; establishes the polar baseline to beat.","marker":"[1]"},{"why":"Defines the polar feature-distortion problem and contributes the GRR attention design; PVP's GRP is contrasted with it and its decomposed attention is adapted.","marker":"[4]"},{"why":"Provides the asymmetric decomposed convolution idea that PD-Conv builds on to separate distortions into plane-wise 2D convolutions.","marker":"[27]"},{"why":"Establishes polar grid partitioning for LiDAR perception, the representational foundation PVP adopts.","marker":"[26]"},{"why":"A camera-based occupancy baseline included in the main comparison table that PVP outperforms.","marker":"[14]"}],"fun_headline_variants":["Polar voxel model beats Cartesian baselines","PVP: Polar grid boosts 3D occupancy accuracy","Polar representation improves 3D semantic occupancy","New polar model tops 3D occupancy benchmarks","Polar voxel predictor wins on OpenOccupancy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements assume the published baseline results are directly comparable; the paper aligns pipelines with the benchmark but does not rerun the baselines or verify identical training schedules, voxelization ranges, and evaluation code.","fun_headline_variants_meta":{"raw":{"variants":["Polar voxel model beats Cartesian baselines","PVP: Polar grid boosts 3D occupancy accuracy","Polar representation improves 3D semantic occupancy","New polar model tops 3D occupancy benchmarks","Polar voxel predictor wins on OpenOccupancy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1681,"prompt_tokens":916,"completion_tokens":765,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":691}},"tokens_in":532,"tokens_out":765,"duration_ms":6864,"temperature":1.0,"reasoning_tokens":691,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:39:36.585202+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun PointOcc and M-CONet with the exact training schedule (25 epochs, batch size 8, V100 GPUs), voxel range [0.3, 73]m on the rho axis, and the same evaluation code; if the mIoU gaps over these baselines shrink to well below the reported 1.9% and 7.9%, the central claim of state-of-the-art performance is refuted.","supporting_citations":[{"cited_title":"Openoccupancy: A large scale benchmark for sur- rounding semantic occupancy perception","cited_arxiv_id":null,"evidence_quote":"Supplies the OpenOccupancy dataset, evaluation protocol, and the CONet baselines (L-CONet, M-CONet, C-CONet) that PVP inherits its pipeline from and compares against."},{"cited_title":"Partner: Level up the polar representa- tion for lidar 3d object detection, 2023","cited_arxiv_id":null,"evidence_quote":"Defines the polar feature-distortion problem and contributes the GRR attention design; PVP's GRP is contrasted with it and its decomposed attention is adapted."},{"cited_title":"Cylindrical and asymmetrical 3d convolution networks for lidar segmenta- tion","cited_arxiv_id":null,"evidence_quote":"Provides the asymmetric decomposed convolution idea that PD-Conv builds on to separate distortions into plane-wise 2D convolutions."},{"cited_title":"Polarnet: An improved grid representation for online lidar point clouds se- mantic segmentation","cited_arxiv_id":null,"evidence_quote":"Establishes polar grid partitioning for LiDAR perception, the representational foundation PVP adopts."}],"review_version":1}